Noise log likelihood recomputed at each likelihood call

Currently calling the likelihood.log_likelihood() during sampling recalculates noise log likelikood at each step:

def noise_log_likelihood(self):
        log_l = 0
        for interferometer in self.interferometers:
            mask = interferometer.frequency_mask
            log_l -= noise_weighted_inner_product(
                interferometer.frequency_domain_strain[mask],
                interferometer.frequency_domain_strain[mask],
                interferometer.power_spectral_density_array[mask],
                self.waveform_generator.duration) / 2
        return float(np.real(log_l))

In most cases it is inconsequential, as it does not invoke calling the waveform generator. But for multi-banded likelihoods, where the data is stored on just a fraction of the frequency points, the time wasted on re-computation can become noticeable (as the noise log likelihood is calculated on unbanded frequency points).

Is there a reason why noise_log_likelihood() is recomputed at each likelihood call, instead of being pre-computed before sampling? If not, I would suggest changing it, as it is a small and easy change to make.